Module neuralops.tests.test_qwen3_5_comprehensive

Subagent 6: Qwen3.5-0.8B Dedicated Test Suite (DeltaNet, GQA Attention, MoE).

Functions

def test_qwen3_5_capability_checks()

Expand source code

def test_qwen3_5_capability_checks():
    is_compat_causal, _ = is_module_declared_compatible("adaptive_span", "qwen3_5")
    assert is_compat_causal is True

    is_compat_svd, _ = is_module_declared_compatible("adaptive_svd", "qwen3_5")
    assert is_compat_svd is True

def test_qwen3_5_family_detection()

Expand source code

def test_qwen3_5_family_detection():
    m = MockQwen3_5FullModel()
    assert detect_family(m) == "qwen3_5"
    adapter = ModelAdapter(m)
    assert adapter.causal is True
    assert adapter.bidirectional is False
    assert adapter.position_embedding == "rotary"

def test_qwen3_5_forward_backward()

Expand source code

def test_qwen3_5_forward_backward():
    m = MockQwen3_5FullModel()
    x = torch.randn(2, 4, 32, requires_grad=True)
    out = m(x)
    assert out.shape == (2, 4, 2)
    assert torch.isfinite(out).all()

    loss = out.sum()
    loss.backward()
    assert x.grad is not None
    assert torch.isfinite(x.grad).all()

def test_qwen3_5_leaf_discovery()

Expand source code

def test_qwen3_5_leaf_discovery():
    m = MockQwen3_5FullModel()
    leaves = get_attention_modules(m)
    assert len(leaves) == 2
    assert any("deltanet" in name for name, _ in leaves)
    assert any("attn" in name for name, _ in leaves)

def test_qwen3_5_lora_dynamic_discovery()

Expand source code

def test_qwen3_5_lora_dynamic_discovery():
    m = MockQwen3_5FullModel()
    targets = find_all_linear_submodule_names(m)
    assert "qk_proj" in targets
    assert "q_proj" in targets
    assert "gate" in targets
    assert "classifier" not in targets  # Head excluded

Classes

class MockQwen3_5DeltaNet (hidden=32, qk_dim=16, v_dim=16)

Expand source code

class MockQwen3_5DeltaNet(nn.Module):
    def __init__(self, hidden=32, qk_dim=16, v_dim=16):
        super().__init__()
        self.qk_proj = nn.Linear(hidden, qk_dim)
        self.v_proj = nn.Linear(hidden, v_dim)
        self.out_proj = nn.Linear(v_dim, hidden)
    def forward(self, x, **kwargs):
        return self.out_proj(self.v_proj(x)), None

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes::

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call :meth:to, etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

:ivar training: Boolean represents whether this module is in training or evaluation mode. :vartype training: bool

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Ancestors

• torch.nn.modules.module.Module

Methods

def forward(self, x, **kwargs) ‑> Callable[..., typing.Any]

Expand source code

def forward(self, x, **kwargs):
    return self.out_proj(self.v_proj(x)), None

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class MockQwen3_5FullModel

Expand source code

class MockQwen3_5FullModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.config = type("Config", (), {
            "model_type": "qwen3_5",
            "hidden_size": 32,
            "num_attention_heads": 2,
            "num_key_value_heads": 1,
            "num_hidden_layers": 2,
        })()
        self.layer0_deltanet = MockQwen3_5DeltaNet()
        self.layer0_moe = MockQwen3_5MoEBlock()
        self.layer1_attn = MockQwen3_5GatedAttention()
        self.layer1_moe = MockQwen3_5MoEBlock()
        self.classifier = nn.Linear(32, 2)

    def forward(self, x):
        h, _ = self.layer0_deltanet(x)
        h = h + self.layer0_moe(h)
        h, _ = self.layer1_attn(h)
        h = h + self.layer1_moe(h)
        return self.classifier(h)

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes::

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call :meth:to, etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

:ivar training: Boolean represents whether this module is in training or evaluation mode. :vartype training: bool

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Ancestors

• torch.nn.modules.module.Module

Methods

def forward(self, x) ‑> Callable[..., typing.Any]

Expand source code

def forward(self, x):
    h, _ = self.layer0_deltanet(x)
    h = h + self.layer0_moe(h)
    h, _ = self.layer1_attn(h)
    h = h + self.layer1_moe(h)
    return self.classifier(h)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class MockQwen3_5GatedAttention (hidden=32, q_dim=32, kv_dim=16)

Expand source code

class MockQwen3_5GatedAttention(nn.Module):
    def __init__(self, hidden=32, q_dim=32, kv_dim=16):
        super().__init__()
        self.q_proj = nn.Linear(hidden, q_dim)
        self.k_proj = nn.Linear(hidden, kv_dim)
        self.v_proj = nn.Linear(hidden, kv_dim)
        self.o_proj = nn.Linear(q_dim, hidden)
    def forward(self, x, **kwargs):
        return self.o_proj(self.q_proj(x)), None

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes::

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call :meth:to, etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

:ivar training: Boolean represents whether this module is in training or evaluation mode. :vartype training: bool

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Ancestors

• torch.nn.modules.module.Module

Methods

def forward(self, x, **kwargs) ‑> Callable[..., typing.Any]

Expand source code

def forward(self, x, **kwargs):
    return self.o_proj(self.q_proj(x)), None

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class MockQwen3_5MoEBlock (hidden=32, num_experts=4, expert_dim=64)

Expand source code

class MockQwen3_5MoEBlock(nn.Module):
    def __init__(self, hidden=32, num_experts=4, expert_dim=64):
        super().__init__()
        self.gate = nn.Linear(hidden, num_experts, bias=False)
        self.experts = nn.ModuleList([
            nn.Sequential(nn.Linear(hidden, expert_dim), nn.SiLU(), nn.Linear(expert_dim, hidden))
            for _ in range(num_experts)
        ])
    def forward(self, x):
        scores = torch.softmax(self.gate(x), dim=-1)
        out = sum(scores[..., i:i+1] * exp(x) for i, exp in enumerate(self.experts))
        return out

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes::

import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call :meth:to, etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

:ivar training: Boolean represents whether this module is in training or evaluation mode. :vartype training: bool

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Ancestors

• torch.nn.modules.module.Module

Methods

def forward(self, x) ‑> Callable[..., typing.Any]

Expand source code

def forward(self, x):
    scores = torch.softmax(self.gate(x), dim=-1)
    out = sum(scores[..., i:i+1] * exp(x) for i, exp in enumerate(self.experts))
    return out

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.